Engine assembly process feature selection method and system

By combining multiple feature selection algorithms, the assembly process features that have the greatest impact on engine vibration are screened out, which solves the problem of difficulty in preventing vibration over-error in the prior art and improves the assembly qualification rate and efficiency.

CN114626465BActive Publication Date: 2024-11-15NANJING UNIV OF AERONAUTICS & ASTRONAUTICS +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210263901.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-17
Publication Date
2024-11-15
Estimated Expiration
2042-03-17

AI Technical Summary

Technical Problem

The prior art is difficult to effectively screen out the key assembly process characteristics that have the greatest impact on engine vibration, which makes it difficult to prevent and solve the vibration error problem during engine assembly, and the debugging process is blind and consumes a lot of manpower and time.

Method used

A combination of multiple feature selection algorithms is adopted, including low variance filtering, mutual information method, Pearson correlation coefficient, Kendall's level coefficient, LGBM weight analysis and SVM importance analysis. Through multiple rounds of screening and fusion, the impact on engine vibration is extracted. The largest assembly process feature.

Benefits of technology

The key assembly process characteristics that have the greatest impact on engine vibration are effectively screened out, which improves the assembly qualification rate, reduces the blindness of commissioning and labor costs, and improves assembly efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114626465B_ABST
    Figure CN114626465B_ABST
Patent Text Reader

Abstract

The present invention discloses an engine assembly process feature selection method and system, which relates to the technical field of engine assembly. The method comprises: obtaining assembly samples generated by engine assembly; performing filtering feature selection on the assembly samples using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set; performing encapsulation feature selection on the first feature set using LGBM weight analysis to obtain a second feature set; performing encapsulation feature selection on the second feature set using SVM importance analysis to obtain a third feature set; the third feature set comprises a plurality of assembly process features that have the greatest impact on engine vibration, which are screened out after encapsulation feature selection is performed on the second feature set. The present invention can effectively screen out the key assembly process features that have the greatest impact on engine vibration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of engine assembly, and in particular to a method and system for selecting process features of engine assembly. Background Art

[0002] Vibration tolerance has always been one of the key issues in the engine assembly production process, which greatly hinders the production progress of the engine. There are many factors that affect the vibration level of the engine. Not only are the magnitudes and properties of various exciting forces in different types of engines different, but the exciting forces of each engine in the same model of engine are also different. These exciting forces mainly come from the following aspects. First, it comes from the imbalance of the rotor during high-speed rotation. The root cause is the residual dynamic balance of the rotor and the influence of the bearing. Secondly, the exciting force generated by the parts, such as the impact of various gear teeth, bearing manufacturing errors, wear or defects, imbalance of the transmission shaft and gears, improper assembly of various rotating parts, tightness, clearance or concentricity tolerance, etc. can all generate exciting forces. Since the vibration level of each engine in the actual assembly work can only be measured after the final assembly is completed, it is impossible to calculate the final vibration value based on the relevant information of each assembly component before assembly. The assembly of each engine is a complex and time-consuming task. If the vibration is out of tolerance after the final assembly is completed, the assembled engine needs to be disassembled, parts replaced or assembly operations adjusted, and then reassembled and tested. Moreover, since it is impossible to clearly identify the key engine assembly process that causes vibration tolerance, each debugging is somewhat blind, consuming a lot of manpower and time. Analyzing the engine assembly data, selecting features from all assembly process features, and extracting the key assembly process features that have the greatest impact on vibration can provide important investigation targets for reference for the assembly personnel's debugging work, guide the correction of the assembly process, and improve the engine assembly qualification rate.

[0003] Feature selection is an important data preprocessing process in machine learning. It can eliminate irrelevant or redundant features, select features that are truly relevant to the label, and assist in understanding the process of data generation. Common feature selection methods include filtering methods and encapsulation methods. The filtering method selects features based on performance metrics, regardless of the algorithm model. It first performs feature selection on the original data set and then trains the model. The filtering method can sort a single feature or evaluate the entire feature subset. It has lower computational cost, faster computing speed and better generalization ability. Its measurement methods can be roughly divided into: information, distance, consistency, similarity and statistical metrics. The encapsulation method uses the algorithm model as part of the selection process to select feature subsets. The model created is regarded as a black box evaluator. Since the encapsulation method requires iterative calculations, the computational cost is higher than that of the filtering method.

[0004] Feature selection in existing research usually only uses one method to screen key features, that is, either a filtering method or an encapsulation method. This cannot avoid ignoring other information of the features and losing some important information due to the different focus of each method. Since it is impossible to comprehensively and objectively screen the features from multiple angles, it is impossible to effectively screen out the key assembly process features that have the greatest impact on engine vibration. Summary of the invention

[0005] The purpose of the present invention is to provide an engine assembly process feature selection method and system, which can effectively screen out key assembly process features that have the greatest impact on engine vibration.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for selecting process features of an engine assembly process, the method comprising:

[0008] Acquire an assembly sample generated by engine assembly; the assembly sample includes all assembly process features during the engine assembly process and vibration values ​​after the engine is assembled;

[0009] Performing filtering feature selection on the assembly sample by using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set; the first feature set includes a plurality of assembly process features screened out after performing filtering feature selection on the assembly sample;

[0010] Performing packaging feature selection on the first feature set using LGBM weight analysis to obtain a second feature set; the second feature set includes a plurality of assembly process features screened out after performing packaging feature selection on the first feature set;

[0011] The second feature set is subjected to packaged feature selection by using SVM importance analysis to obtain a third feature set; the third feature set includes a plurality of assembly process features having the greatest impact on engine vibration that are screened out after the packaged feature selection is performed on the second feature set.

[0012] Optionally, the filtering feature selection of the assembly samples using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain the first feature set specifically includes:

[0013] Sorting the assembly process features in the assembly sample by using low variance filtering to obtain a first sorting result;

[0014] Using a mutual information method to sort the assembly process features in the assembly sample, to obtain a second sorting result;

[0015] sorting the assembly process features in the assembly samples using the Pearson correlation coefficient to obtain a third sorting result;

[0016] Using the Kendall rank coefficient to sort the assembly process features in the assembly sample to obtain a fourth sorting result;

[0017] Using a voting method to merge the first sorting result, the second sorting result, the third sorting result and the fourth sorting result, to obtain a total number of votes for each assembly process feature in the assembly sample;

[0018] Sorting the assembly process features from high to low according to the total number of votes to obtain a fifth sorting result;

[0019] The first feature set is formed by using the first 10% of the assembly process features in the fifth sorting result.

[0020] Optionally, the using LGBM weight analysis to perform encapsulated feature selection on the first feature set to obtain the second feature set specifically includes:

[0021] Acquire a source data set; the source data set includes a plurality of assembly samples generated by assembling a plurality of engines; the assembly samples correspond one to one with the engines;

[0022] According to each of the assembly process features in the first feature set, data is extracted from the source data set to obtain a first training data set; the first training data set includes a plurality of extracted assembly samples; each of the extracted assembly samples only includes each of the assembly process features in the first feature set;

[0023] Using the first training data set to train the LGBM model, to obtain the weight of each of the assembly process features in the first feature set;

[0024] Sorting the assembly process features in the first feature set from large to small according to the weights to obtain a sixth sorting result;

[0025] According to the sixth sorting result, each of the assembly process features in the first feature set is screened by forward search to obtain a second feature set.

[0026] Optionally, the using SVM importance analysis to perform encapsulated feature selection on the second feature set to obtain a third feature set specifically includes:

[0027] According to each of the assembly process features in the second feature set, data is extracted from the source data set to obtain a second training data set; the second training data set includes a plurality of extracted assembly samples; each of the extracted assembly samples only includes each of the assembly process features in the second feature set;

[0028] Using the second training data set to train the SVM model, to obtain a vector corresponding to the second feature set;

[0029] Obtaining data values ​​of each dimension of the vector according to the vector;

[0030] Calculate the importance coefficient of each of the assembly process features in the second feature set according to the data values ​​of each dimension of the vector;

[0031] Sorting the assembly process features in the second feature set from large to small according to the importance coefficients to obtain a seventh sorting result;

[0032] Recording the last assembly process feature in the seventh sorting result;

[0033] The last assembly process feature in the seventh sorting result is removed from the second feature set, the second feature set is updated, and the process returns to the step of "extracting data from the source data set according to each assembly process feature in the second feature set to obtain a second training data set" until the second feature set is empty;

[0034] Sorting the recorded assembly process features in order according to the order of recording to obtain an eighth sorting result;

[0035] The third feature set is formed by using the last 50% of the assembly process features in the eighth sorting result.

[0036] The present invention also provides the following solution:

[0037] An engine assembly process feature selection system, the system comprising:

[0038] An assembly sample acquisition module is used to acquire assembly samples generated by engine assembly; the assembly samples include all assembly process features during the engine assembly process and vibration values ​​after the engine is assembled;

[0039] A filtering feature selection module is used to perform filtering feature selection on the assembly sample by using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set; the first feature set includes a plurality of assembly process features screened out after filtering feature selection is performed on the assembly sample;

[0040] A first packaged feature selection module, used to perform packaged feature selection on the first feature set by using LGBM weight analysis to obtain a second feature set; the second feature set includes a plurality of assembly process features screened out after performing packaged feature selection on the first feature set;

[0041] The second encapsulated feature selection module is used to perform encapsulated feature selection on the second feature set using SVM importance analysis to obtain a third feature set; the third feature set includes multiple assembly process features that have the greatest impact on engine vibration, which are screened out after encapsulated feature selection on the second feature set.

[0042] Optionally, the filtering feature selection module specifically includes:

[0043] A low variance filtering unit, used for sorting the assembly process features in the assembly sample by using low variance filtering to obtain a first sorting result;

[0044] A mutual information method unit, used for sorting the assembly process features in the assembly sample by using the mutual information method to obtain a second sorting result;

[0045] A Pearson correlation coefficient unit, used for sorting the assembly process features in the assembly sample by using the Pearson correlation coefficient to obtain a third sorting result;

[0046] A Kendall rank coefficient unit, used for ranking each of the assembly process features in the assembly sample by using the Kendall rank coefficient to obtain a fourth ranking result;

[0047] A voting method unit, used for fusing the first sorting result, the second sorting result, the third sorting result and the fourth sorting result by using a voting method to obtain a total number of votes for each of the assembly process features in the assembly sample;

[0048] A first sorting unit, configured to sort the assembly process features from high to low according to the total number of votes to obtain a fifth sorting result;

[0049] The first feature set forming unit is used to form a first feature set using the first 10% of the assembly process features in the fifth sorting result.

[0050] Optionally, the first encapsulated feature selection module specifically includes:

[0051] A source data set acquisition unit, used to acquire a source data set; the source data set includes a plurality of assembly samples generated by assembling a plurality of engines; the assembly samples correspond to the engines one by one;

[0052] A first data extraction unit is used to extract data from the source data set according to each of the assembly process features in the first feature set to obtain a first training data set; the first training data set includes a plurality of extracted assembly samples; each of the extracted assembly samples only includes each of the assembly process features in the first feature set;

[0053] An LGBM model training unit, configured to train the LGBM model using the first training data set to obtain a weight of each of the assembly process features in the first feature set;

[0054] A second sorting unit is used to sort the assembly process features in the first feature set from large to small according to the weights to obtain a sixth sorting result;

[0055] The second feature set obtaining unit is used to screen the assembly process features in the first feature set by forward search according to the sixth sorting result to obtain a second feature set.

[0056] Optionally, the second encapsulated feature selection module specifically includes:

[0057] A second data extraction unit is used to extract data from the source data set according to each of the assembly process features in the second feature set to obtain a second training data set; the second training data set includes a plurality of extracted assembly samples; each of the extracted assembly samples only includes each of the assembly process features in the second feature set;

[0058] An SVM model training unit, used to train the SVM model using the second training data set to obtain a vector corresponding to the second feature set;

[0059] A unit for obtaining data values ​​of each dimension, used for obtaining data values ​​of each dimension of the vector according to the vector;

[0060] An importance coefficient calculation unit, used for calculating the importance coefficient of each of the assembly process features in the second feature set according to the data values ​​of each dimension of the vector;

[0061] A third sorting unit is used to sort the assembly process features in the second feature set from large to small according to the importance coefficient to obtain a seventh sorting result;

[0062] A recording unit, used for recording the last assembly process feature in the seventh sorting result;

[0063] A feature elimination and return unit, used for eliminating the last assembly process feature in the seventh sorting result from the second feature set, updating the second feature set and returning to the second data extraction unit until the second feature set is empty;

[0064] A fourth sorting unit, used to sort the recorded assembly process features in sequence according to the order of recording to obtain an eighth sorting result;

[0065] The third feature set forming unit is used to form a third feature set using the last 50% of the assembly process features in the eighth sorting result.

[0066] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0067] The engine assembly process feature selection method and system disclosed in the present invention integrate a variety of feature selection algorithms, including low variance filtering, mutual information method, Pearson correlation coefficient, Kendall rank coefficient method for filtering feature selection, and LGBM weight analysis and SVM importance analysis for encapsulation feature selection, and perform multiple rounds of screening on assembly process features that affect the engine vibration level, and finally extract the key assembly process features that are most relevant to the vibration, that is, the assembly process features that have the greatest impact on the engine vibration, so as to effectively screen out the key assembly process features that have the greatest impact on the engine vibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0069] Figure 1 A flow chart of an embodiment of a method for selecting features of an engine assembly process according to the present invention;

[0070] Figure 2 It is a flow chart of the method for implementing the iterative multi-round integrated engine key assembly process feature selection of the present invention;

[0071] Figure 3 A comparison chart of the learning effects of the prediction model using different feature selection results in the present invention;

[0072] Figure 4 It is a structural diagram of an embodiment of the engine assembly process feature selection system of the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The purpose of the present invention is to provide an engine assembly process feature selection method and system, which can effectively screen out key assembly process features that have the greatest impact on engine vibration.

[0075] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] Figure 1 Flow chart of an embodiment of the engine assembly process feature selection method of the present invention. Figure 1 , the engine assembly process feature selection method includes:

[0077] Step 101: Acquire an assembly sample generated by engine assembly; the assembly sample includes all assembly process features during the engine assembly process and vibration values ​​after the engine is assembled.

[0078] In this embodiment, 65 assembly process data samples generated in the actual engine assembly work are collected, and each sample has 248 assembly process features. In step 101, the assembly data generated in the engine assembly work (for example, when assembling an engine, an assembly sample will be obtained, and this sample includes: rotor mass, nut tightening torque, fan and transmission shaft clearance and other data, a total of 248 types, "rotor mass", "nut tightening torque" and so on are all assembly process features) are collected as the source assembly data set, and the assembly data of each engine is a sample in the source assembly data set, including multiple assembly process feature data (one of which is the vibration value of the engine after assembly), and all process features in the engine assembly process constitute the source feature set.

[0079] Step 102: Perform filtering feature selection on the assembly samples using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set; the first feature set includes multiple assembly process features screened out after filtering feature selection on the assembly samples.

[0080] The step 102 specifically includes:

[0081] The assembly process features in the assembly samples are sorted by using low variance filtering to obtain a first sorting result.

[0082] The mutual information method is used to sort the assembly process features in the assembly samples to obtain a second sorting result.

[0083] The Pearson correlation coefficient is used to rank the assembly process features in the assembly samples to obtain a third ranking result.

[0084] The Kendall rank coefficient is used to rank the assembly process characteristics in the assembly sample to obtain a fourth ranking result.

[0085] The first sorting result, the second sorting result, the third sorting result and the fourth sorting result are merged by using the voting method to obtain the total number of votes for each assembly process feature in the assembly sample.

[0086] The assembly process features are sorted from high to low according to the total number of votes to obtain a fifth sorting result.

[0087] The first feature set is formed by using the first 10% of assembly process features in the fifth sorting result.

[0088] This step 102 mainly performs filtering feature selection, which selects features with obvious data changes and high correlation with sample labels according to the correlation between feature data from the perspective of data. When performing filtering feature selection, this step 102 uses four methods, namely low variance filtering, mutual information method, Pearson correlation coefficient, and Kendall rank coefficient (the four methods of low variance filtering, mutual information method, Pearson correlation coefficient, and Kendall rank coefficient are used to process the source assembly data set at the same time) to perform the following operations on the source assembly data set:

[0089] The source assembly data set is X = {x i |i=1,2,...,n sample}, where x i represents a sample in the source assembly dataset, n sample Represents the number of samples contained in the source assembly data set. Let y = {y i |i=1,2,...,n sample} is the vibration value corresponding to each sample in the source assembly data set x. Let F = {f i |i=1,2,...,n f} is the feature set of the source assembly dataset, that is, the features contained in each sample in the source assembly dataset, where f i Represents a feature of the sample, n f Indicates the number of features contained in the sample.

[0090] Calculate the variance of each feature and sort the features from large to small according to the variance. The larger the variance of a feature, the greater the data variation. For a feature f0, the calculation formula for its variance is: where s 2 f0 represents the variance of feature f0, x i [f0] represents sample x i The data value of feature f0 in , Represents the average value of the feature f0 in all samples of the source assembly dataset, calculated as:

[0091] Calculate the standard mutual information of each feature with other features and vibration values, and sort the features from small to large according to the standard mutual information with other features and from large to small according to the standard mutual information with vibration values. The smaller the standard mutual information with other features, the greater the independence of the feature, and the larger the standard mutual information with vibration values, the greater the correlation between the feature and vibration. For two features f0 and f1, the standard mutual information calculation formula between them is: Where I(X[f0], X[f1]) represents the standard mutual information between features f0 and f1, X[f0] represents the vector of data values ​​of feature f0 of all samples in the source assembly dataset, X[f1] represents the vector of data values ​​of feature f1 of all samples in the source assembly dataset, p(a, b) is the joint distribution probability of variables a and b, and p(a) and p(b) are the marginal distribution probabilities of variables a and b, respectively. Joint distribution probability and marginal distribution probability are both basic concepts in probability theory and mathematical statistics.

[0092] Calculate the Pearson correlation coefficient between each feature and the vibration value, and sort the features from large to small according to the Pearson correlation coefficient. The feature with a larger Pearson correlation coefficient has a greater correlation with the vibration. For a feature f0 and the vibration value y corresponding to all samples in the source assembly data set X, the Pearson correlation coefficient between them is calculated as: in, represents the Pearson correlation coefficient between features f0 and y, cov(X[f0], y) represents the covariance between feature data X[f0] and y, represents the standard deviation of X[f0], σ y represents the standard deviation of y, and E[] represents the expectation of the contents in []. represents the mean value of X[f0], μ y represents the mean of y.

[0093] Calculate the Kendall rank coefficient of each feature and vibration value, and sort the features from large to small according to the Kendall rank coefficient. The larger the Kendall rank coefficient of a feature, the greater its correlation with vibration. For a feature f0 and the vibration value y corresponding to all samples in the source assembly data set X, the Kendall rank coefficient between them is calculated as: in represents the Kendall rank coefficient between features f0 and y, n sample represents the number of samples in the source assembly data set, which is also the number of elements in X[f0] and y. C represents the number of pairs of elements that are consistent in X[f0] and y (two elements are a pair), and D represents the number of pairs of elements that are inconsistent in X[f0] and y. C and D are calculated as follows:

[0094] For X[f0] and y, their i-th elements are represented by X[f0] i and i Indicates that X[f0] and the elements in the corresponding position in y form an element pair set X[f0] y , which contains the following elements (X[f0] i ,y i )(1≤i≤n sample ). When any two elements in the set X[f0]y (X[f0] i ,y i ) and (X[f0] j ,y j ) have the same ranking, that is, when case 1 or case 2 occurs (case 1: X[f0] i >X[f0] j And y i >y j , Case 2: X[f0] i <X[f0] j And y i <y j ), the two elements are considered consistent. When case 3 or case 4 occurs (case 3: X[f0] i >X[f0] j And y i <y j , Case 4: X[f0] i <X[f0] j And y i >y j ), the two elements are considered inconsistent. When case 5 or 6 occurs (case 5: X[f0] i =X[f0] j , case 6: y i =y j ), the two elements are neither consistent nor inconsistent.

[0095] The above operations give the ranking results of features under different methods. There are four methods: low variance filtering, mutual information method, Pearson correlation coefficient, and Kendall rank coefficient. After processing the source assembly data set, each method will get a ranking result. That is, each method will get a ranking result. Finally, there are four ranking results. Next, the voting method is used to merge the different ranking results (the fusion result is the feature with higher votes, that is, the top 10% of the features with the highest votes from high to low). The process is as follows:

[0096] First, different weights w are assigned to each feature selection method according to the degree of attention paid to the data characteristics. i For example, if the variance of feature data is more important, then a relatively larger weight can be set for low variance filtering. Suppose the number of feature selection methods is N, the number of features is M, and the jth feature is ranked in the ith feature selection method as I. ij (The ranking position is a number, i.e. the rank), then the total number of votes for the jth feature is:

[0097]

[0098] Among them, the total number of votes f j is a specific value. The number of votes for each feature is calculated according to the above method. According to the voting results, the features with higher votes (the top 10% of the features with the highest votes) are selected as the first round of key process feature results selected by the filtering feature selection. The set of multiple selected features (the first feature set) is recorded as F key-1 ,Right now Figure 2 The key feature set F in key-1 .

[0099] Step 103: Use LGBM weight analysis to perform packaged feature selection on the first feature set to obtain a second feature set; the second feature set includes multiple assembly process features screened out after performing packaged feature selection on the first feature set.

[0100] The step 103 specifically includes:

[0101] A source data set is obtained; the source data set includes a plurality of assembly samples generated by assembling a plurality of engines; and the assembly samples correspond to the engines one by one.

[0102] According to each assembly process feature in the first feature set, data is extracted from the source data set to obtain a first training data set; the first training data set includes multiple extracted assembly samples; each extracted assembly sample only includes each assembly process feature in the first feature set.

[0103] The LGBM model is trained using the first training data set to obtain the weights of each assembly process feature in the first feature set.

[0104] The assembly process features in the first feature set are sorted from large to small according to the weights to obtain a sixth sorting result.

[0105] According to the sixth sorting result, each assembly process feature in the first feature set is screened by forward search to obtain a second feature set.

[0106] This step 103 mainly performs packaging feature selection - LGBM weight analysis. This step 103 uses the LGBM weight analysis method to further screen the assembly process features for the first round of key process features obtained in step 102. The process is as follows:

[0107] Based on the LGBM algorithm, the data in the source assembly data set that only retains the first round of key process features is trained to obtain the vibration prediction model. For example, if the first round of key process features F key-1 It contains two process features, rotor mass and nut tightening torque. Each sample in the source assembly data set contains data on multiple features, such as rotor mass, nut tightening torque, and the gap between the fan and the transmission shaft. "Only retain the data in the source assembly data set with the first round of key process features" means that each sample in the source assembly data set only retains the data on the two features, rotor mass and nut tightening torque. The source assembly data set can be regarded as a table. Each row of this table is a sample, and each column is a process feature. The size of the table is (n sample *n f ), n sample is the number of samples, n fis the number of features. The vibration prediction model is the result of data training, and the output of the training is the vibration prediction model. The vibration prediction model is the product of data training. It is an algorithm model that can predict whether the vibration level of the engine after assembly is qualified based on the assembly data of the engine. The input of the model is the assembly data of the engine, and the output of the model is whether the vibration level of the engine after assembly is qualified. This embodiment does not specifically use this prediction model, mainly to obtain the weight of the feature after the data training is completed. LGBM is a decision tree model. In the process of training the data, the weight of the feature is measured according to the total number of times the feature is used for segmentation in all decision trees, so as to mine the importance information of the feature (LGBM is an existing algorithm). After the data is trained, according to the feature weight returned by LGBM (the feature weight is based on the engine assembly data, that is, the data in the source assembly data set that only retains the first round of key process features mentioned in the first sentence of this paragraph, and is obtained by training using LGBM. Each feature in the engine assembly data has a different influence on the vibration level after the engine is assembled. The weight of the feature reflects this influence. The role of LGBM is to mine the feature weight from the data) sort the features from large to small. The feature with the larger weight is more important and the ranking position is higher. Next, based on the feature sorting results, the features are further screened using a forward search method. Forward search means incrementally selecting one feature from the remaining unselected features to add to the feature set each time. If the model training effect is better after adding the feature, the feature is retained, otherwise it is discarded.

[0108] Here are the steps:

[0109] 1. Initialize the feature set F′ as the set of features with the largest feature weight (here F′ is the feature set with the largest weight selected according to the weight returned by LGBM in the previous paragraph), and use the data set that only retains the features in F′ to train the vibration prediction model.

[0110] 2. The first round of key process feature set F is obtained in the filtering feature selection key-1 , the feature in F′ is from F key-1 The features selected according to the weights in the previous section. For the first round of key process features except F′, the remaining features f i (F′ is F key-1 A subset of , the remaining features refer to F key-1 The remaining features in F except the features included in F′, that is, F key-1 The features contained in F but not in F′, f i Used to refer to the remaining features), add features f one by one in descending order of weight (the weight is calculated according to the previous paragraph). i to the set F′(f iis a single feature, F′ is a set of multiple features, add feature f i To the set F′ is to add the feature f i Then, the vibration prediction model is retrained. If the accuracy of the model is improved compared with that before adding, F′=F′∪{f i}, otherwise F′ remains unchanged before adding, and continues to add the next feature until the first round of key process feature set F key-1 All features in have been screened.

[0111] 3. Output F′.

[0112] The output F′ after the forward search is completed is used as the second round of key process feature results selected by the encapsulated feature selection-LGBM weight analysis, and the set of multiple selected features (the second feature set) is recorded as F key-2 ,Right now Figure 2 The key feature set F in key-2 .

[0113] Step 104: Perform packaged feature selection on the second feature set using SVM importance analysis to obtain a third feature set; the third feature set includes multiple assembly process features that have the greatest impact on engine vibration, which are screened out after performing packaged feature selection on the second feature set.

[0114] The step 104 specifically includes:

[0115] According to each assembly process feature in the second feature set, data is extracted from the source data set to obtain a second training data set; the second training data set includes multiple extracted assembly samples; each extracted assembly sample only includes each assembly process feature in the second feature set.

[0116] The SVM model is trained using the second training data set to obtain a vector corresponding to the second feature set.

[0117] Get the data value of each dimension of the vector based on the vector.

[0118] The importance coefficient of each assembly process feature in the second feature set is calculated according to the data value of each dimension of the vector.

[0119] The assembly process features in the second feature set are sorted from large to small according to the importance coefficient to obtain a seventh sorting result.

[0120] Record the last assembly process feature in the seventh sorting result.

[0121] The last assembly process feature in the seventh sorting result is removed from the second feature set, the second feature set is updated and the step of "extracting data from the source data set according to each assembly process feature in the second feature set to obtain a second training data set" is returned until the second feature set is empty.

[0122] The recorded assembly process features are sorted in sequence according to the recording order to obtain an eighth sorting result.

[0123] The third feature set is formed by using the last 50% of the assembly process features in the eighth sorting result.

[0124] This step 104 mainly performs encapsulated feature selection - SVM importance analysis. This step 104 uses the SVM importance analysis method to further screen the second round of key process features obtained in step 103. SVM (support vector machine) searches for the optimal classification hyperplane w·x+b=0 (w·x+b=0 is just to express the principle of SVM and has nothing to do with the second round of key process features. SVM is an existing algorithm. The principle is briefly introduced here because w is used below). The solution of w can be obtained as:

[0125]

[0126] Among them, α i is the Lagrange multiplier, x i is the i-th sample data, y i is the label of the ith sample, and N is the number of samples. w·x+b=0 is the representation of the classification hyperplane. The classification hyperplane can be determined by simply finding the parameter w in the equation w·x+b=0. Formula (1) is the solution formula for w.

[0127] The steps of SVM importance analysis are as follows:

[0128] a. Define a new set F′, and initialize the feature set F′ to be the second round of key process feature set selected by the encapsulated feature selection in step 103 - LGBM weight analysis. The elements contained in F′ are F key-2 Initialize the third round of key process feature set F key-3 Step 104 initializes the feature set F′ based on the key feature set output in step 103 .

[0129] b. Based on the SVM algorithm, the data in the source assembly data set that only retains the process features in F' is trained to obtain the vibration prediction model. When training the data, SVM will calculate w in formula (1). For the i-th feature of the data set (the source assembly data set that only retains the process features in F'), the importance coefficient of the feature is calculated as follows:

[0130] Rc (i)=||w (i) || 2

[0131] Among them, R c (i) represents the importance coefficient of the i-th feature, w (i) represents the data value of the i-th dimension of w. w is a vector. w is calculated by formula (1). The purpose of calculating w is to calculate R c (i) Sort the features according to their importance coefficient from large to small. The smaller the importance coefficient, the lower the importance of the feature.

[0132] c. Add the features at the end of the sorting in step b to F key-3 The end of F′ is deleted and the cycle continues until F′ is empty. This step c adds the features to F one by one in order of importance from small to large. key-3 (Add to the end), for example, if at the beginning F key-3 is an empty set, and feature a is added in the first loop, so F key-3 = {a}, the second loop adds feature b, then F key-3 = {a, b}, feature b is added after feature a, and feature c is added in the third cycle, so F key-3 ={a, b, c}, feature c is added after feature b. key-3 The elements in are in order, and the importance of features increases from the front to the back, that is, the importance of feature a < the importance of feature b < the importance of feature c.

[0133] d. Output F key-3 .

[0134] The output of the above steps is F key-3 That is, the feature set arranged from small to large importance obtained through SVM importance analysis, update F key-3 To keep only F key-3 The features with greater importance in the latter part (the last 50% of the features in order of importance) are selected as the third round of key process features (the third feature set), that is, Figure 2 The key feature set F in key-3 Taking the example in step c above, keep F key-3 The features with greater importance in the latter are the ones that retain F key-3 The position in is the following features, namely feature b and feature c.

[0135] Through the filtering feature selection in step 102, the encapsulation feature selection in step 103 - LGBM weight analysis and the encapsulation feature selection in step 104 - SVM importance analysis, the third round of key process features F key-3 , which is also the key engine assembly process feature that is most relevant to vibration screened out by the present invention. It can provide an important investigation target for reference for the debugging work of engine assemblers, guide the correction of assembly process, and improve the engine assembly qualification rate. At the same time, it can also perform feature dimensionality reduction when constructing the engine vibration prediction model to reduce the model's dependence on data volume.

[0136] This embodiment carried out three rounds of feature screening using the above feature selection method, and screened out 26, 11, and 8 key process features from 248 process features, respectively. In order to verify the effectiveness of the key process features, based on the linear discriminant analysis algorithm, different key process features were used to build an engine vibration prediction model based on the existing assembly data, and the learning effect of the model was compared. The model evaluation index is mainly AUC (area under the ROC curve), supplemented by Accuracy, Precision, and F1-score (a comprehensive indicator based on precision and recall). The test results of the experiment are as follows: Figure 3 As shown, Figure 3 Parts (a), (b), (c), and (d) respectively show the AUC comparison, F1-score comparison, Accuracy comparison, and Precision comparison of prediction models using different numbers of key features in 10 random tests. Through iterative multi-round integrated feature selection, four sets of feature sets were obtained through different feature selection operations. Figure 3 The comparison of the learning effect of the prediction model using different numbers of key features in 10 random tests is recorded. Figure 3It can be seen that for different evaluation indicators, the overall performance of the model training using the key process features after the third round of feature selection is better than that of the other three groups, and the broken line formed by the results of 10 tests is at the top as a whole. With the deepening of feature selection, the learning effect of the model becomes better in turn. Table 1 records the average test results of model training using different key features. The source assembly data set contains 248 process features. Through three rounds of feature selection, 26, 11, and 8 key process features are selected respectively, and the feature compression rate gradually increases to 90%, 96%, and 97%, respectively. It can be seen that with multiple rounds of feature selection, the size of the key feature subset, that is, the feature dimension, gradually decreases, and the score values ​​of various evaluation indicators of the model increase in turn. Compared with the source feature set, the key feature subset after the third round of screening increases the AUC by nearly 8 percentage points, which can reach 95%, and the improvement of F1-score, Accuracy, and Precision is more obvious. Experiments show that the iterative multi-round feature selection method of the present invention can effectively screen out key assembly process features.

[0137] Table 1 Comparison of average learning effects of prediction models using different feature selection results

[0138]

[0139] The engine assembly process feature selection method disclosed in the present invention relates to the fields of machine learning, data analysis, feature engineering, and engine assembly, and is an iterative multi-round integrated engine key assembly process feature selection implementation method, such as Figure 2As shown. Based on the fact that the existing single feature selection method cannot comprehensively and objectively evaluate the features from multiple angles, and thus some important information will be lost, the present invention proposes an iterative multi-round integrated engine key assembly process feature selection method, which integrates multiple feature selection algorithms, including low variance filtering, mutual information method, Pearson correlation coefficient, Kendall rank coefficient method of filtering feature selection, and LGBM weight analysis and SVM importance analysis of encapsulation feature selection, and performs multiple rounds of screening on the assembly process features that affect the engine vibration level, and finally extracts the key assembly process features that are most relevant to the vibration. The present invention takes into account that different methods have different concerns when selecting features. The filtering method is more based on the correlation of data, and the encapsulation method is more based on the contribution of features to the model. The principles and emphases of different feature selection methods are different. Referring to the idea of ​​ensemble learning, different methods are used to perform multiple rounds of screening on the assembly process features of the engine, starting from multiple angles, integrating multiple feature selection methods, and performing feature selection in all the assembly process features of the engine, screening out the key assembly process features that have the greatest impact on the engine vibration, providing important investigation targets for reference for the debugging work of assembly personnel, guiding the correction of assembly process, and improving the qualified rate of engine assembly. At the same time, feature dimensionality reduction can also be performed when building an engine vibration prediction model to reduce the model's dependence on data volume.

[0140] The present invention is based on the fact that existing research technicians only focus on civil or commercial engines. Such engines can obtain a large amount of data in practice or through simulation software. Therefore, only one feature selection method is needed to meet the requirements of feature selection (such as the compression rate of feature dimension). However, for engines for specific purposes with extremely little engine assembly data and particularly many process characteristics, technicians have never used feature selection in previous studies on engine assembly data for specific purposes. A feature selection method that can be applied to the field of engine assembly for specific purposes is provided, and the feature selection technology in machine learning is used for the analysis of engine assembly data for specific purposes. Since the engine assembly data for specific purposes is extremely little and the process characteristics are numerous, a variety of feature selection methods are used in combination to perform multiple rounds of screening of key features to meet the requirements of feature selection.

[0141] Figure 4 This is a structural diagram of an embodiment of the engine assembly process feature selection system of the present invention. Figure 4 , the engine assembly process feature selection system includes:

[0142] The assembly sample acquisition module 401 is used to acquire assembly samples generated by engine assembly; the assembly samples include all assembly process features during the engine assembly process and vibration values ​​after the engine is assembled.

[0143] The filtering feature selection module 402 is used to perform filtering feature selection on the assembly samples using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set; the first feature set includes multiple assembly process features screened out after filtering feature selection on the assembly samples.

[0144] The filtering feature selection module specifically includes:

[0145] The low variance filtering unit is used to sort the assembly process features in the assembly sample by using low variance filtering to obtain a first sorting result.

[0146] The mutual information method unit is used to sort the assembly process features in the assembly sample by using the mutual information method to obtain a second sorting result.

[0147] The Pearson correlation coefficient unit is used to sort the assembly process features in the assembly sample using the Pearson correlation coefficient to obtain a third sorting result.

[0148] The Kendall rank coefficient unit is used to sort the assembly process features in the assembly sample by using the Kendall rank coefficient to obtain a fourth sorting result.

[0149] The voting method unit is used to merge the first sorting result, the second sorting result, the third sorting result and the fourth sorting result by using the voting method to obtain the total number of votes for each assembly process feature in the assembly sample.

[0150] The first sorting unit is used to sort the assembly process features from high to low according to the total number of votes to obtain a fifth sorting result.

[0151] The first feature set forming unit is used to form a first feature set using the first 10% of assembly process features in the fifth sorting result.

[0152] The first packaging feature selection module 403 is used to perform packaging feature selection on the first feature set using LGBM weight analysis to obtain a second feature set; the second feature set includes multiple assembly process features screened out after performing packaging feature selection on the first feature set.

[0153] The first encapsulated feature selection module specifically includes:

[0154] The source data set acquisition unit is used to acquire the source data set; the source data set includes a plurality of assembly samples generated by assembling a plurality of engines; the assembly samples correspond to the engines one by one.

[0155] The first data extraction unit is used to extract data from the source data set according to the assembly process features in the first feature set to obtain a first training data set; the first training data set includes multiple extracted assembly samples; each extracted assembly sample only includes the assembly process features in the first feature set.

[0156] The LGBM model training unit is used to train the LGBM model using the first training data set to obtain the weight of each assembly process feature in the first feature set.

[0157] The second sorting unit is used to sort the assembly process features in the first feature set from large to small according to the weight to obtain a sixth sorting result.

[0158] The second feature set obtaining unit is used to screen the assembly process features in the first feature set by forward search according to the sixth sorting result to obtain the second feature set.

[0159] The second encapsulated feature selection module 404 is used to perform encapsulated feature selection on the second feature set using SVM importance analysis to obtain a third feature set; the third feature set includes multiple assembly process features that have the greatest impact on engine vibration, which are screened out after encapsulated feature selection on the second feature set.

[0160] The second encapsulated feature selection module specifically includes:

[0161] The second data extraction unit is used to extract data from the source data set according to the assembly process features in the second feature set to obtain a second training data set; the second training data set includes multiple extracted assembly samples; each extracted assembly sample only includes the assembly process features in the second feature set.

[0162] The SVM model training unit is used to train the SVM model using the second training data set to obtain a vector corresponding to the second feature set.

[0163] The unit for obtaining the data value of each dimension is used to obtain the data value of each dimension of the vector according to the vector.

[0164] The importance coefficient calculation unit is used to calculate the importance coefficient of each assembly process feature in the second feature set according to the data value of each dimension of the vector.

[0165] The third sorting unit is used to sort the assembly process features in the second feature set from large to small according to the importance coefficient to obtain a seventh sorting result.

[0166] The recording unit is used to record the last assembly process feature in the seventh sorting result.

[0167] The feature elimination and return unit is used to eliminate the last assembly process feature in the seventh sorting result from the second feature set, update the second feature set and return to the second data extraction unit until the second feature set is empty.

[0168] The fourth sorting unit is used to sort the recorded assembly process features in sequence according to the recording order to obtain an eighth sorting result.

[0169] The third feature set forming unit is used to form a third feature set using the last 50% of the assembly process features in the eighth sorting result.

[0170] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0171] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for selecting engine assembly process features, characterized in that: The method comprises: Acquire an assembly sample generated by engine assembly; the assembly sample includes all assembly process features during the engine assembly process and vibration values ​​after the engine is assembled; The assembly sample is subjected to filtering feature selection by using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set; the first feature set includes a plurality of assembly process features screened out after filtering feature selection is performed on the assembly sample; the filtering feature selection is performed on the assembly sample by using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set, specifically including: sorting each of the assembly process features in the assembly sample by using low variance filtering to obtain a first sorting result; sorting each of the assembly process features in the assembly sample by using mutual information method to obtain a second sorting result; sorting each of the assembly process features in the assembly sample by using Pearson correlation coefficient The assembly process features are sorted to obtain a third sorting result; the assembly process features in the assembly sample are sorted using the Kendall rank coefficient to obtain a fourth sorting result; the first sorting result, the second sorting result, the third sorting result and the fourth sorting result are merged using a voting method to obtain a total number of votes for each assembly process feature in the assembly sample; the assembly process features are sorted from high to low according to the total number of votes to obtain a fifth sorting result; the first 10% of the assembly process features in the fifth sorting result are used to form a first feature set; the four methods of low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient are used to process the source assembly data set simultaneously; The first feature set is subjected to encapsulated feature selection by using LGBM weight analysis to obtain a second feature set; the second feature set includes a plurality of assembly process features selected after the first feature set is subjected to encapsulated feature selection; the encapsulated feature selection by using LGBM weight analysis to obtain the second feature set specifically includes: obtaining a source data set; the source data set includes a plurality of assembly samples generated by assembling a plurality of engines; the assembly samples correspond one to one with the engines; according to each of the assembly process features in the first feature set, data extraction is performed on the source data set to obtain a first training data set; the first training data set includes a plurality of extracted assembly samples; each of the extracted assembly samples only includes each of the assembly process features in the first feature set; the LGBM model is trained by using the first training data set to obtain the weight of each of the assembly process features in the first feature set; the assembly process features in the first feature set are sorted from large to small according to the weight to obtain a sixth sorting result; according to the sixth sorting result, the assembly process features in the first feature set are screened by using forward search to obtain the second feature set; the role of LGBM is to mine feature weights from data; The second feature set is subjected to packaged feature selection by using SVM importance analysis to obtain a third feature set; the third feature set includes a plurality of assembly process features having the greatest impact on engine vibration that are screened out after the packaged feature selection is performed on the second feature set.

2. The engine assembly process feature selection method according to claim 1, characterized in that: The encapsulated feature selection is performed on the second feature set by using SVM importance analysis to obtain a third feature set, specifically including: According to each of the assembly process features in the second feature set, data is extracted from the source data set to obtain a second training data set; the second training data set includes a plurality of extracted assembly samples; each of the extracted assembly samples only includes each of the assembly process features in the second feature set; Using the second training data set to train the SVM model, to obtain a vector corresponding to the second feature set; Obtaining data values ​​of each dimension of the vector according to the vector; Calculate the importance coefficient of each of the assembly process features in the second feature set according to the data values ​​of each dimension of the vector; Sorting the assembly process features in the second feature set from large to small according to the importance coefficients to obtain a seventh sorting result; Recording the last assembly process feature in the seventh sorting result; The last assembly process feature in the seventh sorting result is removed from the second feature set, the second feature set is updated and the process returns to step "performing data extraction on the source data set according to each assembly process feature in the second feature set to obtain a second training data set" until the second feature set is empty; Sorting the recorded assembly process features in order according to the order of recording to obtain an eighth sorting result; The third feature set is formed by using the last 50% of the assembly process features in the eighth sorting result.

3. An engine assembly process feature selection system, characterized in that: The system comprises: An assembly sample acquisition module is used to acquire assembly samples generated by engine assembly; the assembly samples include all assembly process features during the engine assembly process and vibration values ​​after the engine is assembled; A filtering feature selection module is used to perform filtering feature selection on the assembly sample by using low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient to obtain a first feature set; the first feature set includes multiple assembly process features screened out after filtering feature selection on the assembly sample; the filtering feature selection module specifically includes: a low variance filtering unit, used to sort each of the assembly process features in the assembly sample by using low variance filtering to obtain a first sorting result; a mutual information method unit, used to sort each of the assembly process features in the assembly sample by using the mutual information method to obtain a second sorting result; a Pearson correlation coefficient unit, used to sort each of the assembly process features in the assembly sample by using the Pearson correlation coefficient to obtain a third sorting result ; Kendall rank coefficient unit, used for sorting each of the assembly process features in the assembly sample using the Kendall rank coefficient to obtain a fourth sorting result; a voting method unit, used for fusing the first sorting result, the second sorting result, the third sorting result and the fourth sorting result using the voting method to obtain the total number of votes for each of the assembly process features in the assembly sample; a first sorting unit, used for sorting the assembly process features from high to low according to the total number of votes to obtain a fifth sorting result; a first feature set forming unit, used for forming a first feature set using the first 10% of the assembly process features in the fifth sorting result; the four methods of low variance filtering, mutual information method, Pearson correlation coefficient and Kendall rank coefficient are simultaneously used to process the source assembly data set; The first encapsulated feature selection module is used to perform encapsulated feature selection on the first feature set using LGBM weight analysis to obtain a second feature set; the second feature set includes multiple assembly process features screened out after performing encapsulated feature selection on the first feature set; the first encapsulated feature selection module specifically includes: a source data set acquisition unit, used to acquire a source data set; the source data set includes multiple assembly samples generated by assembling multiple engines; the assembly samples correspond one to one with the engines; a first data extraction unit, used to extract data from the source data set according to each of the assembly process features in the first feature set, to obtain a first training data set; the first training data set includes A plurality of extracted assembly samples; each of the extracted assembly samples only includes the assembly process features in the first feature set; an LGBM model training unit, used to train the LGBM model using the first training data set to obtain the weights of the assembly process features in the first feature set; a second sorting unit, used to sort the assembly process features in the first feature set from large to small according to the weights to obtain a sixth sorting result; a second feature set obtaining unit, used to screen the assembly process features in the first feature set using forward search according to the sixth sorting result to obtain a second feature set; the role of LGBM is to mine feature weights from data; The second encapsulated feature selection module is used to perform encapsulated feature selection on the second feature set using SVM importance analysis to obtain a third feature set; the third feature set includes multiple assembly process features that have the greatest impact on engine vibration, which are screened out after encapsulated feature selection on the second feature set.

4. The engine assembly process feature selection system according to claim 3, characterized in that: The second encapsulated feature selection module specifically includes: A second data extraction unit is used to extract data from the source data set according to each of the assembly process features in the second feature set to obtain a second training data set; the second training data set includes a plurality of extracted assembly samples; each of the extracted assembly samples only includes each of the assembly process features in the second feature set; An SVM model training unit, used to train the SVM model using the second training data set to obtain a vector corresponding to the second feature set; A unit for obtaining data values ​​of each dimension, used for obtaining data values ​​of each dimension of the vector according to the vector; An importance coefficient calculation unit, used for calculating the importance coefficient of each of the assembly process features in the second feature set according to the data values ​​of each dimension of the vector; A third sorting unit is used to sort the assembly process features in the second feature set from large to small according to the importance coefficient to obtain a seventh sorting result; A recording unit, used for recording the last assembly process feature in the seventh sorting result; A feature elimination and return unit, used for eliminating the last assembly process feature in the seventh sorting result from the second feature set, updating the second feature set and returning it to the second data extraction unit until the second feature set is empty; A fourth sorting unit, used to sort the recorded assembly process features in sequence according to the order of recording to obtain an eighth sorting result; The third feature set forming unit is used to form a third feature set using the last 50% of the assembly process features in the eighth sorting result.